Source-linked AI summary
LANTERN: Language Model Assessment on Noisy and Transformed Tasks for Understanding Error and Robustness Nuances
Vamsi Krishna Kodavali, Rituraj Singh
TL;DR
The paper addresses limited robustness evidence for LLMs under perturbed inputs. It systematically evaluates synthetic perturbations across MCQ and instruction-following tasks, finding robustness varies by perturbation and benchmark, while larger models are not consistently more robust.
Problem
Static evaluation does not fully capture LLM sensitivity to noisy inputs, modified choices, and variable instructions.
Method
The study constructs synthetic and augmented MCQ and instruction-following datasets and evaluates models across perturbation types, scales, and tuning variants.
Results
Robustness varies by perturbation and benchmark; minor typographical errors cause moderate accuracy drops, while replacing correct choices with None of these causes significant degradation.
Takeaways & Limitations
Larger models do not consistently demonstrate higher robustness, so scale alone does not ensure stability under perturbed conditions.
Takeaways & Limitations
The study does not examine synonym or antonym substitutions, broader semantic shifts, or large open-ended sample sets.
Abstract
from arXiv · showhide
Robustness evaluation of large language models (LLMs) remains a critical challenge, particularly in assessing their sensitivity to perturbations in input data. In this work, we systematically evaluate LLM robustness across multiple dimensions, including word error rate, character repetition and duplication, modifications in choices, and variability in instruction following. To facilitate this evaluation, we construct a synthetic and augmented dataset encompassing a diverse set of LLM benchmarks, specifically targeting multiple-choice question (MCQ) datasets and instruction-following tasks. We conduct extensive experiments on LLMs of varying scales-small, medium, and large-as well as across base and instruction-tuned variants. Our analysis quantifies the variability in model responses under perturbed conditions and highlights discrepancies relative to baseline models. The findings provide insights into the stability of LLMs across different evaluation scenarios contributing to the development of more robust and reliable language models as well as robust evaluation methodologies.
1 Introduction
The paper argues that static benchmarks inadequately assess modern LLMs and introduces a robustness evaluation spanning noisy inputs, choice modifications, and instruction-following variability. It constructs augmented datasets and compares diverse model scales and tuning variants against baselines.
- Static benchmarks and earlier NLP metrics are insufficient for evaluating modern LLM reasoning, factual consistency, and adaptability.
- Data leakage and naturally occurring textual errors can undermine evaluation consistency and generalizability.
- The study examines how option choices, question modifications, and instruction-following capabilities affect LLM evaluation.
- The authors evaluate robustness across character-level changes, word error rate, choice modifications, and instruction-following variability.
- A synthetic and augmented dataset covers multiple-choice and instruction-following benchmarks.
- Experiments span small, medium, and large models, including base and instruction-tuned variants, with performance compared against baselines.
2 Related Work
Prior work studies LLM evaluation, option-order sensitivity, and robustness, but commonly used MCQ options and broader perturbations remain insufficiently covered. This work extends evaluation across questions, answer choices, and verifiable instructions.
- Related work organizes prior research around LLMs, evaluation, and robustness on adversarial or noisy tasks.
- Prior studies report sensitivity to example arrangement, answer position, and token-level preferences in LLM outputs.
- Existing studies do not account for common MCQ options such as None of these, All of the above, or Both A and B.
- The paper broadens prior evaluation to perturbations in questions, answer choices, and verifiable instructions across models and tasks.
3 Method
The method evaluates robustness in structured answer selection and open-ended generation by applying controlled perturbations to questions, choices, and instructions. Question perturbations vary in type and intensity, while choice perturbations replace or reorder options and preserve one correct answer.
- The study evaluates structured answer selection and open-ended generation under perturbations.
- Structured answer selection: Structured-answer robustness measures whether models consistently select the correct predefined choice after input rephrasing, noise, or adversarial modification.
- Structured answer selection: For MCQ data D = {(q, C, y, S)}, perturbation function P modifies questions or choices to produce transformed datasets D′ or D′′.
- Questions perturbation: Question perturbations target character- and word-level changes to test answer selection under controlled input variation.
- Questions perturbation: Typographical errors alter randomly selected characters, while letter duplication and removal introduce redundant or missing characters.
- Questions perturbation: Each question perturbation type is applied to 10%, 30%, or 50% of randomly selected words, yielding nine synthetic datasets across type-level combinations.
- Choices perturbation: Choice perturbations replace or permute options using phrases such as None of These, All of the Above, or Both A and B while retaining exactly one correct choice.
- Open-ended generation: Open-ended evaluation represents each sample with a prompt, instruction identifiers, and verifiable functions that check compliance with each instruction.
4 Experiments
The experiments evaluate structured answer selection and open-ended generation under multiple perturbations across benchmarks and model types. Results show that robustness varies by model, dataset, perturbation, and tuning, with choice-structure changes and noisy inputs often reducing performance.
- Evaluation setup: The evaluation covers MMLU, MMLU-Pro, GPQA, and MUSR for structured answer selection, plus IfEval for open-ended generation.
- Evaluation setup: The study compares open-source models from 2B to 13B parameters, including base and instruction-tuned variants.
- Evaluation setup: Perturbation experiments compare original benchmark scores with modified datasets using zero-shot evaluation, except for five-shot MMLU-Pro.
- Question perturbations: Typographical errors reduce MMLU accuracy consistently as perturbation rates increase, with Phi-3.5-mini-Instruct dropping 10.6%, while GPQA shows no clear trend.
- Question perturbations: Letter duplication and letter removal generally degrade performance, although Gemma-2-2B variants remain relatively robust on MMLU-Pro.
- Choice perturbations: Replacing incorrect choices with “None of these” often slightly improves accuracy, whereas removing correct choices or adding multiple-choice dependencies substantially degrades performance.
- Open-ended generation: IfEval performance declines 2.85% under perturbations; instruction-tuned models are more resilient, while base LLaMA models show approximately 4% average accuracy drops.
5 Discussion and Analysis
The study tests whether perturbations preserve dataset distributions and semantic structure, finding controlled changes with stable statistical properties and reproducible results.
- Distributional analysis: The two-sample non-parametric K-S test compares original and perturbed dataset distributions after text preprocessing.The preprocessing removes special characters and stopwords, then applies lemmatization before computing word-frequency distributions.
- Visualization: Figure 6 visualizes the distributions of original and perturbed data and their log-scaled differences.The figure accompanies the distributional comparison and embedding analysis.
- Distributional analysis: 0.9750 p-value for MMLU-Pro indicates no significant statistical difference between original and perturbed word-frequency distributions.The reported threshold is p ≥ 0.05.
- Semantic analysis: Minimal embedding divergence indicates that perturbations preserve the original questions' overall statistical and semantic properties.The comparison uses PCA projections of embeddings for original and perturbed questions.
- Reproducibility: Average standard deviations of 0.47, 0.48, and 1.01 across three MMLU random seeds indicate stable observations across perturbation runs.The datasets were regenerated with three different random seeds to assess variability.
6 Conclusion
The paper evaluates LLM robustness under multiple perturbation types and benchmarks, finding that robustness varies by perturbation and benchmark rather than following model scale alone.
- Findings: LLMs show varying robustness across typographical errors, choice modifications, and instruction variability in multiple-choice and instruction-following tasks.Minor typographical errors cause moderate accuracy drops, while replacing correct choices with “None of these” causes significant degradation.
- Findings: Larger models do not consistently demonstrate higher robustness under perturbed conditions.The conclusion states that scale alone does not confer stability.
- Implications: The findings support evaluation methodologies and training strategies that account for real-world input variability.Future work proposes adaptive fine-tuning and adversarial training for structured perturbations.
Limitations
The study leaves several robustness dimensions and evaluation scales unexplored, including semantic substitutions, models larger than 14B, and larger open-ended samples.
- Perturbation scope: Synonym and antonym substitutions and broader semantic shifts in word distributions are outside the study’s perturbation analysis.The authors leave these modifications for future research.
- Model scope: Models larger than 14B are not included, leaving the effect of increased parameter counts on robustness unresolved.The authors note that larger-scale models may exhibit different generalization behaviors.
- Sampling scope: The open-ended generation analysis uses a limited number of samples, restricting deeper assessment of output consistency and reliability.The authors propose expanding the sample set in future work.
A.1 IfEval samples
The appendix presents IfEval examples illustrating prompts, verifiable instructions, and introduced sample changes.
- Example input: The appendix includes an example input taken from the IfEval paper.The example is presented as background for the dataset.
- IfEval visualization: Figure 7 shows a prompt with verifiable instructions displayed in italic.The caption directs readers to the IfEval paper for additional information.
- Perturbation examples: Figure 8 presents samples containing the changes introduced by the study.The figure caption identifies these as examples of the perturbations.
A.3 Summarized results for Modified Choices
This section compiles results for modified-choice evaluations across MMLU, MMLU-PRO, GPQA, and MUSR. It includes benchmark figures across multiple variants and score tables for modified questions, choices, and narrations.
- Benchmark coverage: Results are reported for modified choices across MMLU, MMLU-PRO, GPQA, and MUSR benchmarks.The figures cover v0, v1, and v2 variants, while the tables separately report modified-choice scores for these benchmarks.